GenomeFlowInterface Settings:
shuffle: True
pos_sampling_rate: 0.05
validation_chroms: ['chr22']
holdout_chroms: ['chr1', 'chr8', 'chr21']
optimizer: adam
learning rate: 0.0003
batch size: 256
epoch size: 2500000
max num of epochs: 100
early stopping metrics: auPRC
early stopping patience: 4

Epoch 1:
Metrics across all datasets:
Balanced Accuracy: 95.27%	auROC: 0.997	 auPRC: 0.573
Recall at 5% | 10% | 25% | 50% FDR: 0.9% | 4.3% | 31.2% | 63.0%	Num Positives: 1415	 Num Negatives: 996701


New best auPRC. Saving model.


Epoch 2:
Metrics across all datasets:
Balanced Accuracy: 94.68%	auROC: 0.995	 auPRC: 0.640
Recall at 5% | 10% | 25% | 50% FDR: 0.1% | 14.2% | 47.1% | 72.3%	Num Positives: 1415	 Num Negatives: 996701


New best auPRC. Saving model.


Epoch 3:
Metrics across all datasets:
Balanced Accuracy: 95.97%	auROC: 0.996	 auPRC: 0.618
Recall at 5% | 10% | 25% | 50% FDR: 0.0% | 0.0% | 44.5% | 70.0%	Num Positives: 1415	 Num Negatives: 996701



Epoch 4:
Metrics across all datasets:
Balanced Accuracy: 94.79%	auROC: 0.996	 auPRC: 0.632
Recall at 5% | 10% | 25% | 50% FDR: 1.1% | 2.0% | 42.9% | 72.2%	Num Positives: 1415	 Num Negatives: 996701



Epoch 5:
Metrics across all datasets:
Balanced Accuracy: 92.63%	auROC: 0.994	 auPRC: 0.626
Recall at 5% | 10% | 25% | 50% FDR: 0.4% | 16.8% | 44.2% | 70.2%	Num Positives: 1415	 Num Negatives: 996701



Epoch 6:
Metrics across all datasets:
Balanced Accuracy: 93.00%	auROC: 0.994	 auPRC: 0.627
Recall at 5% | 10% | 25% | 50% FDR: 0.1% | 17.6% | 42.5% | 71.3%	Num Positives: 1415	 Num Negatives: 996701



Epoch 7:
Metrics across all datasets:
Balanced Accuracy: 93.16%	auROC: 0.990	 auPRC: 0.632
Recall at 5% | 10% | 25% | 50% FDR: 0.4% | 18.7% | 48.8% | 72.7%	Num Positives: 1415	 Num Negatives: 996701


Finished training after 7 epochs.
The best model's architecture and weights (from epoch 2) were saved to /users/marinovg/2018-01-25-histone-mods/tfdragonn-SequenceAndDnaseClassifier-K562-DNAse-Duke-cuts-CTCFL/model.arch.json and /users/marinovg/2018-01-25-histone-mods/tfdragonn-SequenceAndDnaseClassifier-K562-DNAse-Duke-cuts-CTCFL/model.weights.h5